Are there limitations as to what I can set the grid size of a CUDA kernel to be? I ran into a problem where kernels were not launching with a grid size of 33 x 33 but were able to launch when the grid size was 32 x 32. Is there any reason for this to occur? Or is it likely that changing the number of blocks from 32 x 32 to 33 x 33 broke some other constraint?

dim3 blockSize(8, 8);
dim3 gridSize(32, 32);

cudaDeviceSynchronize();
set_start<<<gridSize, blockSize>>>(some_params);

The above works.

dim3 blockSize(8, 8);
dim3 gridSize(33, 33);

cudaDeviceSynchronize();
set_start<<<gridSize, blockSize>>>(some_params);

The above does not work.

kernel & main:

__global__
void set_start(double * const H , double * const HU , double * const HV , 
           double * const E , const int Na)
{
int j = threadIdx.x + blockIdx.x*blockDim.x + 1;
int i = threadIdx.y + blockIdx.y*blockDim.y + 1;

if(i >= Na-1 || j >= Na-1)
    return;

H[i*Na+j]  = 1.0 + exp(-100.0*((E[j-1]-0.75)*(E[j-1]-0.75)+(E[i-1]-0.75)*(E[i-1]-0.75))) + 0.5*exp(-100.0*((E[j-1]-0.75)*(E[j-1]-0.75)+(E[i-1]-0.25)*(E[i-1]-0.25)));
HU[i*Na+j] = 0; 
HV[i*Na+j] = 0;
}

int main(int argc, char** argv){

double* E_d;
cudaMalloc(&E_d, sizeof(double) * (Nh+1));
set_E<<<64, (Nh/64) + 1>>>(E_d, dx, Nh);

int Na = 259;
double *H_d, *HU_d, *HV_d, *Ht_d, *HUt_d, *HVt_d;

cudaMalloc(&H_d , sizeof(double) * Na * Na);
cudaMalloc(&HU_d, sizeof(double) * Na * Na);
cudaMalloc(&HV_d, sizeof(double) * Na * Na);

dim3 blockSize(8, 8);
//dim3 gridSize(((Na-1)/blockSize.x) + 1, ((Na-1)/blockSize.y) + 1);
//dim3 gridSize(33, 33);
dim3 gridSize(32, 32);

cudaDeviceSynchronize();
set_start<<<blockSize, gridSize>>>(H_d, HU_d, HV_d, E_d, Na);
}

This was on CUDA 7.0.